Google is forming a new team dedicated to developing artificial intelligence models capable of simulating the physical world.
The team will be led by Tim Brooks, a former co-lead on OpenAI’s video generation model, Sora, who joined Google DeepMind’s AI research lab in October. Brooks announced the initiative in a post on X on Monday, emphasizing the ambitious goals of the project.
“DeepMind has ambitious plans to create massive generative models that simulate the world,” Brooks said. “I’m hiring for a new team with this mission.”
The new team will build upon the work of Google’s existing Gemini, Veo, and Genie teams, aiming to address key challenges and scale models to unprecedented computational levels. Gemini, Google’s flagship AI model, handles tasks such as image analysis and text generation, while Veo focuses on video generation.
Genie, a key component of the project, is Google’s approach to “world models,” AI capable of simulating real-time 3D environments and interactive games. The latest version of Genie, previewed in December, can generate a wide range of playable 3D worlds.
Brooks explained that scaling AI training using video and multimodal data is considered essential for achieving artificial general intelligence, which refers to AI capable of performing any task a human can.
The goal is for world models to power domains such as visual reasoning, real-time simulation, planning for autonomous agents, and interactive entertainment.
The new team will focus on real-time interactive generation tools, integrating their models with existing multimodal models like Gemini to advance these capabilities.
World models are gaining significant attention from both startups and large tech companies, including AI researcher Fei-Fei Lee’s World Labs, Decart, and Odyssey. These models are seen as potentially revolutionizing fields like interactive media, video games, movies, and simulations for robotics training.
Some startups, like Odyssey, have committed to collaborating with creative professionals rather than replacing them. It remains to be seen whether Google will take a similar approach.
Additionally, there are unresolved copyright concerns related to the training of world models, particularly if they use video game content or other unlicensed material. Google, which owns YouTube, maintains that it has permission to use YouTube videos for model training, but has not disclosed which specific videos are being sourced.
